Effects of Racial Bias in Dermatology Datasets on Convolutional Neural Networks

Abstract: 

Artificial Intelligence (AI) has shown significant promise in medical diagnosis, especially in dermatology for early detection of skin cancer. Racial disparities in the medical field have been firmly established, especially in dermatology, but have not yet been rectified. This deficiency of representation extends to AI systems, which often perform poorly on dark skin tones due to a biased training dataset dominated by light skin images. This project investigates how bias in training data can affect AI's ability to detect skin cancer accurately. The project examines this by training two models: one on the lighter-skin-heavy HAM10000 dataset and another on the more diverse DDI dataset. Model performance is evaluated in various racial subgroups to detect inequalities in sensitivity, specificity and overall accuracy. Through cross-evaluation, this study examines how racial imbalance in training data affects diagnostic accuracy. The goal is to better understand how imbalanced data can lead to unequal outcomes, and to highlight the importance of using diverse datasets in building fair and effective AI tools in dermatology.